Try the Work

Inspiration

In 2023, I wrote Harmony around a human question: how can people and intelligent tools grow together without losing agency? Ongoing Harmony 2 work sharpened that into a practical rule—keep human judgment visible, carry evidence with claims, and make learning lead to real capability instead of another answer. Try the Work applies that philosophy to one of life's highest-stakes transitions: getting ready for a job.

What it does

Paste an authorized public or synthetic job description. Try the Work turns it into a realistic, AI-augmented mini-shift. The learner performs observable work, uses AI the way people increasingly will on the job, and receives a candidate-owned Work Receipt with source citations, scored evidence, improvement recommendations, and explicit untested gaps.

For candidates, it is a safer way to practice before applying and show what they can actually do. For hiring teams, it is a foundation for fairer, job-relevant tryouts. For employers, the same structure supports uptraining and cross-training without pretending one exercise measures a whole person.

How it works

1. Read only the authorized role source.
2. Separate supported requirements from assumptions.
3. Build a bounded mini-shift with realistic decisions and artifacts.
4. Let the learner work with deterministic guidance and optional localhost-only Ollama assistance.
5. Check observable evidence—not personality—and name what remains untested.
6. Require human review before exporting a learner-owned receipt or a deeper Codex work packet.

The included renewal-rescue shift runs without an API key or paid model. Optional Ollama support stays on numeric loopback, receives no tools, and cannot take outside action. Codex is the deeper build-and-improve path, while the learner remains the decision-maker.

How I used Codex and GPT-5.6

GPT-5.6 through Codex was a material engineering collaborator: researching the problem, challenging novelty, shaping the source-bounded architecture, generating and revising implementation, building adversarial tests, reviewing security and fairness boundaries, and turning repeated human feedback into a clearer product and a real-motion demo. I retained all product, identity, claims, approval, and release decisions.

Challenges

The hard part was avoiding resume theater. A polished artifact can still overstate ability. The product therefore separates observed evidence from untested duties, rejects unsupported commitments, records provenance, and keeps hiring decisions human. A second challenge was delivering useful AI assistance without creating a billing or privacy trap, so the verified path is deterministic and the optional local model is bounded.

Accomplishments

• A working end-to-end mini-shift, not a slide deck.
• Candidate-owned evidence with exact source citations and an explicit untested gap.
• Human-approved export and zero silent actions.
• 17 focused tests covering state, citations, math, prompt-like source text, unsupported claims, hashes, approval, local lessons, and runtime-mode contracts.
• A reproducible local build and optional offline-assistance path.

What I learned

AI belongs inside modern job rehearsal because using it well is itself a skill. But AI should not become the hidden judge. The most trustworthy experience shows what was observed, what source governs, what remains unknown, and what the human can do next.

What's next

Job-profile evidence links, employer-authored rehearsal packs, more job families, accessibility expansion, longitudinal cross-training, and deeper learner-approved Codex workflows—always with provenance, explicit limits, and human control.

Built With

• typescript